TONY XU
About me
Hello! I'm a third-year PhD candidate at the Department of Medical Biophysics at the University of Toronto supervised by Dr. Anne Martel and Dr. Maged Goubran. My research focuses on improving self-supervised methods applied to huge medical images to reduce the reliance of deep learning methods on expert-generated annotations. I'm also broadly interested in using deep learning in healthcare to accelerate and enhance medical workflows. I'm honored to have recently received the Google PhD Fellowship in health and bioscience.
Before my PhD, I completed my undergraduate studies at UBC with a major in Electrical Engineering specializing in Biomedical Engineering and a minor in Physics. I'm an avid space enthusiast, enjoy learning and building new things, playing badminton, and going to the gym in my free time. Thanks for visiting my website!
Publications
Detecting noisy labels with repeated cross-validations
Jianan Chen*, Vishwesh Ramanathan*, Tony Xu, Anne Martel
Medical Image Computing and Computer Assisted Intervention, 2024
Interpretable Assessment of Lymph Nodes in Specimens Post-Neoadjuvant Systemic Therapy Using AI
Dina Bassiouny, Tony Xu, Chetan Srinidhi, Michael Sze Wai Lam, Sharon Nofech-Mozes, Anne Martel
United States and Canadian Academy of Pathology Annual Meeting, 2024
Masked Image Modeling for Label-Efficient Segmentation in Two-Photon Excitation Microscopy
Tony Xu, Matthew Rozak, Edward Ntiri, Adrienne Dorr, James Mester, Bojana Stefanovic, Anne Martel, Maged Goubran
MICCAI Workshop in Medical Image Learning with Limited and Noisy Data, 2023
Resource and data efficient self supervised learning
Ozan Ciga*, Tony Xu*, Anne L. Martel
arXiv preprint, 2021
Self supervised contrastive learning for digital histopathology
Ozan Ciga, Tony Xu, Anne L. Martel
Machine Learning with Applications, 2022
Overcoming the limitations of patch-based learning to detect cancer in whole slide images
Ozan Ciga, Tony Xu, Sharon Nofech-Mozes, Shawna Noy, Fang-I Lu, Anne L. Martel
Scientific Reports, 2021
Stuff I've done
🧠SELMA3D Challenge
SELMA3D is a challenge on using self-supervised learning to improve the segmentation of structures inside 3D light-sheet microscopy images. We developed a novel self-supervised method and segmentation pipeline to tackle the challenge, more details on this coming soon! Our method ended up obtaining third place on the challenge, and we presented our work at a Medical Image Computing and Computer Assisted Intervention (MICCAI) 2024 workshop.
; SemiCOL Challenge
SemiCOL is a computational challenge using self-supervised learning approaches to tackle colorectal cancer detection. The pipeline we developed for the challenge used label-efficient methods to pretrain feature extractors (PAWS and DINO), that were later applied to a tissue segmentation task. We used both convolution-based (UNet) and Transformer-based (UNETR) segmentation models. The algorithm ended up obtaining second place in both arms of the challenge, and I presented the algorithm during the 19th European Congress on Digital Pathology (EDCP 2023)!
🧲DDPM for MRI Denoising
A project I made for an AI course during my PhD. Diffusion models are typically used as generative models that recover high-quality images from random noise in an iterative manner. However, if you give them an image halfway through this iterative process, they can also be used for denoising! I applied this idea and extended the original DDPM model into 3D to denoise MR images. I was motivated to do this project because I was very interested in learning more about DDPMs and how they work at a deeper level.
🐚Samsung Research America - Think Tank Team
I worked for a summer as a computational scientist intern with a fantastic team situated in Mountain View, California. With TTT, I ideated, experimented, and created technologies to push the frontier of existing Samsung products, and sometimes invent new ones! I probably learned something new every day I spent on the team, whether it be technical (like SLaM) or otherwise (like surfing).
❄Machine Learning Approaches to Accelerate Cryo-EM
My capstone project for my Major in Electrical Engineering in collaboration with Gandeeva Therapeutics! Cryo-EM is a new technology capable of creating detailed 3D maps of proteins with high precision. Using traditional and deep-learning computer vision techniques, I helped create a pipeline to automate the imaging process of the cryo-EM microscope.
🔬Multi-Resolution Vision Transformer for Subtype Classification in Ovarian Cancer Whole-Slide Histopathology Images
My thesis for my Major in Electrical Engineering in collaboration with the AIM Lab at UBC! Whole-slide images contain a wealth of data, but their immense size (~100,000 x 100,000 pixels) makes them difficult to process. Usually, computer vision algorithms will comb through the image patch-by-patch (~224 x 224 pixels), but combining information from each patch in the slide is nontrivial. I developed a method to extract, and combine information from patches at multiple resolutions using the new Vision Transformer model.
🔭Variable Source Detection in the CHIME Telescope
The thesis I completed for my Minor in Physics in collaboration with CHIME! The CHIME Telescope monitors a huge swath of the sky every single day, giving it the unique ability to track random transient events. I first cleaned and reduced the size of the ringmap data, and then developed an automated algorithm to detect legitimate sources varying over a few days.
🔬Undergraduate Research Student at Sunnybrook Research Institute
I spent a summer (and some more time remotely) researching with the Sunnybrook Research Institute in Toronto! I worked on two projects during this time. The first one describes a way to improve the detection of the extent of breast cancer in whole-slide histopathology images. The second explores the application of a new wave of self-supervised machine learning algorithms (like SimCLR) to a variety of histopathology datasets and tasks.
🌴Amazon
I spent a summer as an SDE Intern at Amazon, where I worked with a number of talented individuals on the eCommerce Financial Integrations (eCFI) team on financial data aggregation and tracking. I helped design and build an operational dashboard website to consolidate and securely manage the workflows my team tracked. Apart from learning about a variety of AWS services, I had my first experience with writing, documenting, and reviewing code to be used for an industry-level application.
🃏Slay the Spire Neural Card Recommender
Slay the Spire is a roguelike deckbuilder game where you ascend through spire floors while battling enemies with a deck you build along the way. I created a recommender system using a transformer-based encoder model to generate a pair of embedding vectors. One embedding vector encodes a card choice, and the second encodes your general gamestate (takes into account deck and relics). The model essentially uses the distance between these vectors to recommend card choices as you progress through the game. Still a work in progress, keep an eye out for the link soon!
💥Baryon Acoustic Oscillations and CHIME
A poster I created for an introductory cosmology course I took for my physics minor. In the early stages of the universe's development, protons, photons, and electrons were tightly coupled in a primordial fluid. This fluid underwent acoustic oscillations (equivalent of a sound wave in the fluid), that left imprints on the distribution of matter on huge scales. Investigations on this length scale can lead to better constraints on the evolution of the universe, and my poster uses simulated hydrogen density data to demonstrate this.
🧠Neureka 2020 - EEG Seizure Detection
Neureka was an international challenge hosted by Novela Neurotech and NeuroTechX on seizure prediction using EEG voltage data. A Short-Time Fourier Transform and other denoising methods were applied to convert time series signal into a heatmap of frequency over time. By using a simple pretrained ResNet, we determined the approximate start and end of seizures in scalp EEG, and our method obtained 4th place.
💪FlexAI
I worked with a small team of wonderful individuals at FlexAI as a machine learning engineering intern for 8 months. My main contributions were towards detecting fine-grained form errors in real world exercise videos to provide insight on a user's exercise performance. I researched and implemented numerous cutting-edge approaches used for 3D (2D + time) video data and pose detection.
💭MINT
I am the lead of the ML sub-team of Multifaceted Innovations in NeuroTechnology (MINT). We are working on trying to create a home-automation app that is controlled using the mind. I built a data aquisition and preprocessing pipeline, and a simple classification model for EEG data that determines a person's motion intent.
🌏Exchange at Nanyang Technological University
I went on exhange at NTU in Singapore. It was fascinating to experience the multitude of different cultures in and around Singapore, and meeting new people was fun too. And yes, the food is in fact as good as they say.
🌌VLBI Image Reconstruction
I was inspired by the imaging of the black hole M87 to apply machine learning to reconstruct images from incomplete UV-plane data. The further away an object is, the larger the apeture needed to see it. Interferometry links many telescopes across the world to piece together an incomplete picture of a distant object.
📚Famous Author Natural Language Generation
My final project for the machine learning course that I took during exchange at NTU. I investigate increasingly complex language models on generating a text in the style of a famous historical author, and contrast the improvements that each model brings. I learned a lot about the evolution of language-related algorithms in ML, and gained experience with their properties.
🐶Doggin' Dog GAN
A project using generative adversarial networks to generate fake dog images from the Stanford Dogs dataset. We also tried using an encoding model to apply funky GAN-like features to a real dog image.
🛑SmartStop
Our submission for the Decode Congestion hackathon. We used a Raspberry Pi with a camera module to control a blinking 4 or 3-way stop sign, improving the intuitiveness and safety of the intersection. I coded the stop sign controller and created a segmentation model to detect vehicles and pedestrians in real-time using the Berkeley DeepDrive dataset.